hex.deeplearning

Class DeepLearningTask

java.lang.Object

jsr166y.ForkJoinTask<java.lang.Void>

jsr166y.CountedCompleter

water.H2O.H2OCountedCompleter<T>

water.DTask<T>

water.MRTask<T>

hex.FrameTask<DeepLearningTask>

hex.deeplearning.DeepLearningTask

All Implemented Interfaces:

java.io.Serializable, java.lang.Cloneable, java.util.concurrent.Future<java.lang.Void>, jsr166y.ForkJoinPool.ManagedBlocker, water.Freezable

public class DeepLearningTask
extends FrameTask<DeepLearningTask>

See Also:

Serialized Form

Nested Class Summary

Nested classes/interfaces inherited from class hex.FrameTask

FrameTask.ExtractDenseRow

Nested classes/interfaces inherited from class water.DTask

water.DTask.DKeyTask<T extends water.DTask.DKeyTask,V extends water.Keyed>, water.DTask.RemoveCall

Field Summary

Fields inherited from class hex.FrameTask

_dinfo, _jobKey, _shuffle, _useFraction

Fields inherited from class water.MRTask

_appendables, _fr, _fs, _hi, _keys, _left, _lo, _nhi, _nleft, _nlo, _nrite, _rite, _run_local, _topLocal

Fields inherited from class water.DTask

_ex, _modifiesInputs

Constructor Summary

Constructors

Constructor and Description
DeepLearningTask(water.Key jobKey, DeepLearningModelInfo inputModel, float fraction, int iteration) The only constructor

Method Summary

Methods

Modifier and Type	Method and Description
void	applyMiniBatchUpdate(int n) Apply the gradient to update the weights
static void	applyModelUpdates(Neurons[] neurons) Helper to apply back-propagation without clearing out the gradients afterwards Used for gradient checking
protected void	chunkDone(long n) After each chunk, add the number of processed rows to the counter
protected boolean	chunkInit() Override this to initialize at the beginning of chunk processing.
protected int	getMiniBatchSize() Return the mini-batch size Note: If this is overridden, then applyMiniBatch must be overridden as well to perform the model/weight mini-batch update
static Neurons[]	makeNeuronsForTesting(DeepLearningModelInfo minfo)
static Neurons[]	makeNeuronsForTraining(DeepLearningModelInfo minfo)
DeepLearningModelInfo	model_info() Accessor to the object containing the (final) state of the Deep Learning model Should only be queried after calling this.doAll(Frame training)
protected void	postGlobal() After all reduces are done, the driver node calls this method to clean up This is only needed if we're not inside a DeepLearningTask2 (which will do the reduction between replicated data workers).
protected void	postLocal() After all maps are done on a node, this is called to store the per-node model into DKV (for elastic averaging) Otherwise, do nothing.
void	processRow(long seed, DataInfo.Row r) Process one training row at a time (online learning)
void	reduce(DeepLearningTask other) Average the per-node models (for elastic averaging, already wrote them to DKV in postLocal()) This is a no-op between F/J worker threads (operate on the same weights/biases)
protected void	setupLocal() Transfer ownership from global (shared) model to local model which will be worked on
static void	step(long seed, Neurons[] neurons, DeepLearningModelInfo minfo, DeepLearningModelInfo consensus_minfo, boolean training, double[] responses, double offset) Forward propagation assumption: layer 0 has _a filled with (horizontalized categoricals) double values

Methods inherited from class hex.FrameTask

closeLocal, dinfo, map, processRow

Methods inherited from class water.MRTask

appendables, asyncExec, asyncExec, asyncExec, asyncExec, asyncExecOnAllNodes, block, compute2, dfork, dfork, dfork, dfork, dinvoke, doAll, doAll, doAll, doAll, doAll, doAll, doAll, doAll, doAll, doAllNodes, getResult, isReleasable, map, map, map, map, map, map, map, map, map, map, map, map, onCompletion, onExceptionalCompletion, outputFrame, outputFrame, outputFrame, priority, profString, self, setProfile

Methods inherited from class water.DTask

copyOver, getDException, hasException, logVerbose, onAck, onAckAck, setException

Methods inherited from class water.H2O.H2OCountedCompleter

clone, compute, frozenType, icer, nextThrPriority, read_impl, read, readJSON_impl, readJSON, write_impl, write, writeJSON_impl, writeJSON

Methods inherited from class jsr166y.CountedCompleter

addToPendingCount, compareAndSetPendingCount, complete, exec, getCompleter, getPendingCount, getRawResult, setCompleter, setPendingCount, setRawResult, tryComplete

Methods inherited from class jsr166y.ForkJoinTask

adapt, adapt, adapt, cancel, compareAndSetForkJoinTaskTag, completeExceptionally, fork, get, get, getException, getForkJoinTaskTag, getPool, getQueuedTaskCount, getSurplusQueuedTaskCount, helpQuiesce, inForkJoinPool, invoke, invokeAll, invokeAll, invokeAll, isCancelled, isCompletedAbnormally, isCompletedNormally, isDone, join, peekNextLocalTask, pollNextLocalTask, pollTask, quietlyComplete, quietlyInvoke, quietlyJoin, reinitialize, setForkJoinTaskTag, tryUnfork

Methods inherited from class java.lang.Object

equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

DeepLearningTask

public DeepLearningTask(water.Key jobKey,
                DeepLearningModelInfo inputModel,
                float fraction,
                int iteration)

The only constructor

Parameters:

jobKey -

inputModel - Initial model state

fraction - Fraction of rows of the training to train with

iteration -

Method Detail

model_info

public final DeepLearningModelInfo model_info()

Accessor to the object containing the (final) state of the Deep Learning model Should only be queried after calling this.doAll(Frame training)

Returns:

"The" final model after one Map/Reduce iteration

setupLocal

protected void setupLocal()

Transfer ownership from global (shared) model to local model which will be worked on

Overrides:

setupLocal in class FrameTask<DeepLearningTask>

chunkInit

protected boolean chunkInit()

Description copied from class: FrameTask

Override this to initialize at the beginning of chunk processing.

Overrides:

chunkInit in class FrameTask<DeepLearningTask>

Returns:

whether or not to process this chunk

processRow

public final void processRow(long seed,
              DataInfo.Row r)

Process one training row at a time (online learning)

Overrides:

processRow in class FrameTask<DeepLearningTask>

Parameters:

seed - Seed is only used if reproducible mode is enabled

r - Row (must be dense for now)

applyMiniBatchUpdate

public void applyMiniBatchUpdate(int n)

Apply the gradient to update the weights

Overrides:

applyMiniBatchUpdate in class FrameTask<DeepLearningTask>

Parameters:

n - number of trained examples in this last mini batch

applyModelUpdates

public static void applyModelUpdates(Neurons[] neurons)

Helper to apply back-propagation without clearing out the gradients afterwards Used for gradient checking

Parameters:

neurons -

getMiniBatchSize

protected int getMiniBatchSize()

Description copied from class: FrameTask

Return the mini-batch size Note: If this is overridden, then applyMiniBatch must be overridden as well to perform the model/weight mini-batch update

Overrides:

getMiniBatchSize in class FrameTask<DeepLearningTask>

Returns:

chunkDone

protected void chunkDone(long n)

After each chunk, add the number of processed rows to the counter

Overrides:

chunkDone in class FrameTask<DeepLearningTask>

Parameters:

n - Number of processed rows

postLocal

protected void postLocal()

After all maps are done on a node, this is called to store the per-node model into DKV (for elastic averaging) Otherwise, do nothing.

Overrides:

postLocal in class water.MRTask<DeepLearningTask>

reduce

public void reduce(DeepLearningTask other)

Average the per-node models (for elastic averaging, already wrote them to DKV in postLocal()) This is a no-op between F/J worker threads (operate on the same weights/biases)

Overrides:

reduce in class water.MRTask<DeepLearningTask>

Parameters:

other -

postGlobal

protected void postGlobal()

After all reduces are done, the driver node calls this method to clean up This is only needed if we're not inside a DeepLearningTask2 (which will do the reduction between replicated data workers). So if replication is disabled, and every node works on partial data, then we have work to do here (model averaging).

Overrides:

postGlobal in class water.MRTask<DeepLearningTask>

makeNeuronsForTraining

public static Neurons[] makeNeuronsForTraining(DeepLearningModelInfo minfo)

makeNeuronsForTesting

public static Neurons[] makeNeuronsForTesting(DeepLearningModelInfo minfo)

step

public static void step(long seed,
        Neurons[] neurons,
        DeepLearningModelInfo minfo,
        DeepLearningModelInfo consensus_minfo,
        boolean training,
        double[] responses,
        double offset)

Forward propagation assumption: layer 0 has _a filled with (horizontalized categoricals) double values

Parameters:

seed -

neurons -

minfo -

consensus_minfo -

training -

responses - Standardized response(s)